Effects of climate and land-use changes on fish catches across lakes at a global scale

HAL Id: hal-02647929

https://hal.inrae.fr/hal-02647929

Submitted on 8 Jun 2021

HAL is a multi-disciplinary open access archive for the deposit and dissemination of sci- entific research documents, whether they are pub- lished or not. The documents may come from teaching and research institutions in France or abroad, or from public or private research centers.

L’archive ouverte pluridisciplinaire HAL, est destinée au dépôt et à la diffusion de documents scientifiques de niveau recherche, publiés ou non, émanant des établissements d’enseignement et de recherche français ou étrangers, des laboratoires publics ou privés.

Distributed under a Creative Commons Attribution| 4.0 International License

across lakes at a global scale

Yu-Chun Kao, Mark Rogers, David Bunnell, Ian Cowx, Song Qian, Orlane Anneville, T. Douglas Beard, Alexander Brinker, J. Robert Britton, René

Chura-Cruz, et al.

To cite this version:

Yu-Chun Kao, Mark Rogers, David Bunnell, Ian Cowx, Song Qian, et al.. Effects of climate and land-

use changes on fish catches across lakes at a global scale. Nature Communications, Nature Publishing

Group, 2020, 11 (1), �10.1038/s41467-020-14624-2�. �hal-02647929�

Effects of climate and land-use changes on fi sh catches across lakes at a global scale

Yu-Chun Kao et al.

Globally, our knowledge on lake fi sheries is still limited despite their importance to food security and livelihoods. Here we show that fish catches can respond either positively or negatively to climate and land-use changes, by analyzing time-series data (1970–2014) for 31 lakes across five continents. We find that effects of a climate or land-use driver (e.g., air temperature) on lake environment could be relatively consistent in directions, but consequential changes in a lake-environmental factor (e.g., water temperature) could result in either increases or decreases in fi sh catch in a given lake. A subsequent correlation analysis indicates that reductions in fi sh catch was less likely to occur in response to potential climate and land-use changes if a lake is located in a region with greater access to clean water.

This fi nding suggests that adequate investments for water-quality protection and water-use ef fi ciency can provide additional bene fi ts to lake fi sheries and food security.

https://doi.org/10.1038/s41467-020-14624-2

OPEN

#A full list of authors and their affiliations appears at the end of the paper.

1234567890():,;

L akes, natural and man-made, are important reserves of accessible freshwater and are frequently managed for pro- visioning ecosystem services

, such as drinking water, irrigation, hydropower generation, and fisheries

. Consequently, biota in lake ecosystems could be subject to stressors associated with human activities, such as water withdrawal, floodplain development, exotic species introduction, and overexploitation

. Two important sources of anthropogenic stressors are climate and land-use changes, which may alter a lake ecosystem directly through changing water temperature (WT) and water level (WL), and indirectly through the balance between inputs and outputs of nutrients and sediments, as well as changes in food-web dynamics

.

Among lake ecosystem services, fisheries are especially vul- nerable to environmental changes because fish are ectothermic and fish distribution is usually limited by structural

, thermal

, and chemical

barriers. Environmental changes driven by climate and land-use changes have been linked to major shifts in fish catches (CATCHs) and species composition in many lakes around the world, such as lakes Erie

, Kinneret

, Naivasha

, Peipsi

, and Victoria

, as well as the collapse of Aral Sea fish- eries

. Reduced lake CATCHs caused by climate and land-use changes can threaten food security and livelihoods of millions of people worldwide, especially in impoverished countries where rural poor communities may not have appropriate alternative sources of animal protein and employment opportunity

.

Due to their protracted and compounding nature, under- standing the effects of climate and land-use changes on lake CATCHs requires long-term data, which may include climate and land-use data at a spatial scale of the whole catchment (or drai- nage basin) and environmental, biological, and fishery data at a spatial scale of the whole lake. This intensive data requirement has restricted assessments of how lake environment and lake CATCHs respond to climate and land-use changes to lakes in a few data-rich regions, mostly located in North America and Europe

. As a result, given the large number (>1.43 million) of lakes

and their diversity in size, productivity, species compo- sition, and fisheries

, our current understanding of how changing climate and land-use influences lake CATCHs is very limited at a global scale and is generally lacking for lakes in Africa and Asia where CATCHs are the largest (Supplementary Fig. 1).

The goal of this study is to understand how climate and land- use changes affect lake CATCHs at a global scale. We use a

Bayesian networks modeling approach

to analyze time-series data for the 31 study lakes (Fig. 1) over the period 1970–2014. We choose this modeling approach because it can minimize negative effects of data limitations on the statistical power of our analysis, by accounting for our prior understanding of multi-level, causal processes that could underlie changes in CATCHs in response to changes in climate and land use across lakes. Our Bayesian net- works model (BNM) represents the hypothesized causal processes by expressing them as conditional probabilistic relationships between model variables at multiple levels: between climate and land use (operating at a catchment scale) and lake environ- ment

, and between lake environment and (total) fish bio- mass

and thus (total) CATCH

, as shown in Fig. 2. Due to the regional difference in data availability, our study lakes and study period are selected to ensure that we can include approximately equal numbers of study lakes in Africa, the Americas, Asia, and Europe and in both tropic and temperate regions (Fig. 1). The study lakes are diverse in terms of both socio-economic environment and hydrogeomorphology (Sup- plementary Data 2). However, most study lakes in Europe and the Americas have a better socio-economic environment than the study lakes in Asia and Africa. The diversity of these lakes allows us to identify the characteristics of lakes where CATCHs can be vulnerable to climate and land-use changes, which has implica- tions for identifying vulnerable lake fisheries around the globe.

We conduct three analyses in this study. First, we determine how the climate and land-use drivers affect CATCHs across the 31 study lakes in 1970–2014 by estimating coefficients of our BNM. Second, by implementing the resulting BNM, we run simulations to assess the change in each of climate and land-use drivers associated with a (simulated) 25% decrease in CATCH from the 1970–2014 median (hereafter, a 25% catch decrease) in each lake. Third, we conduct a correlation analysis to identify whether any socio-economic characteristic(s) associated with the catchment or hydrogeomorphological characteristic(s) of the lake corresponded to the magnitudes of changes in any of the climate and land-use drivers that were associated with a 25% catch decrease. To improve clarity, Fig. 3 presents a flowchart that summarizes the procedures of these three analyses and Table 1 provides a list of symbols used in the development of BNM.

We demonstrate that climate and land-use changes can have strong effects on CATCHs across lakes at a global scale, but the effects can be either positive or negative. The subsequent

60°N

30°N

0°

30°S

60°S

120°W 60°W 0° 60°E 120°E

Fig. 1 Global distribution of the 31 study lakes.The map was generated by using R⁵⁸and the R package rworldmap⁶²to plot the centers of study lakes (as given in Supplementary Data 2) on a base world map, of which layers, including country borders, continent outlines, global lakes, and latitude–longitude grids, were obtained from a public domain map dataset Natural Earth (www.naturalearthdata.com).

correlation analysis shows that a lake located in a region with greater access to clean water less likely experiences a substantial decrease in CATCH in response to potential climate and land-use changes. These results suggest that investments made for water- quality protection and water-use efficiency can also provide benefits to lake fisheries and food security.

Results

How climate and land-use changes affected CATCHs. The BNM showed that climate and land-use changes could result in either increases or decreases in CATCHs across the 31 study lakes (Table 2). These results accounted for fish stocking and fishing effort (EFF), which both could have strong effects on CATCHs

(Table 2). Based on the 75% one-tailed credible intervals (CIs) of BNM coefficients, the effects of a climate or land-use driver (one of air temperature (AT), precipitation (PRE),

and agricultural land use (LUag) in this study) on lake- environmental factors (WT, WL, and primary productivity) were more consistent in directions (positive or negative) across lakes than the consequential changes in a lake-environmental factor on CATCHs across lakes. Specifically, four of the six BNM variable pairs between the climate land-use drivers and lake- environmental factors (brown arrows in Fig. 2; Table 2) were consistent in the directional relationships across at least half of the 31 study lakes: AT was positively related to WT in all 31 study lakes, but negatively related to WL in 17 lakes, PRE was positively related to WL in 29 lakes, and LUag was positively related to chlorophyll a (CHL), the measure of primary productivity in this study, in 16 lakes. In contrast, none of the 5 BNM variable pairs relating lake-environmental factors to each other or to CATCH (blue arrows in Fig. 2; Table 2) had a consistent directional relationship across the study lakes, although the closest was WT being negatively related to CHL in 15 lakes.

Air temperature

Water

temperature Water level

Chlorophyll a

Fish catch

Fish stocking

Fishing effort

Precipitation Agricultural

land use

Fig. 2 Hypothesized causal processes represented in the Bayesian networks model.Variables associated with catchment climate and land use are in oval-brown boxes. Variables associated with lake environment are in blue-rectangular boxes. Variables associated withfish catch,fish stocking, andfishing effort are in orange-hexagon boxes.

How climate and land-use changes affected fish

catches

Changes in drivers associated with a 25% catch decrease

Characteristics of lakes where fish catches are

vulnerable Gather data and

derive BNM inputs

Derive probability distributions of

drivers

Derive prior distributions of BNM coefficients

Derive posterior distributions of BNM coefficients

(Table 2)

Draw random samples for BNM coefficients

Draw random samples for

each driver

Derive distributions of each driver and its effects on lake-environmental variables associated with a 25%

catch decrease (Figs. 4–6)

Conduct a correlation analysis between the median magnitude of changes for each driver and each socio-economic

and hydrogeomorphological characteristic (Fig. 7) Estimate the median magnitude of changes for each driver associated

with a 25% catch decrease Gather data and derive inputs

for socio-economic and hydrogeomorphological characteristics

Fig. 3 Summaryflowchart for the procedures of three analyses conducted in this study.BNMBayesian networks model.

Changes in drivers associated with a 25% catch decrease. Our simulations showed that a 25% catch decrease (from the 1970–2014 medians) was associated with both colder and warmer AT across the 31 study lakes (Fig. 4a), and the effects of AT on CATCHs were mostly mediated though changes in WT (Fig. 4e, h) rather than WL (Fig. 4f) or CHL (Fig. 4g). For 8 of the 31 lakes, a simulated 25%

catch decrease was associated with colder AT (i.e., a positive rela- tionship), with medians of 0.9–3.6 °C colder than the observed 1970–2014 medians (Fig. 4a). For another eight lakes, however, a simulated 25% catch decrease was associated with warmer AT (i.e., a negative relationship), with medians of 0.8–1.7 °C warmer than the observed 1970–2014 medians. For these 16 lakes, the simulated relationship between AT and WT was consistently

positive (Fig. 4b), whereas the simulated relationships between AT and WL (Fig. 4c) and between AT and CHL (Fig. 4d) could be either positive, negative, or mixed. Translating these simulated changes to effects on CATCHs revealed that the effects of WT, resulting from changes in AT, were important in 8 of these 16 lakes (i.e., leading to a >9.1% catch decrease in simulations; Fig. 4e), even though both colder and warmer WT could lead to a catch decrease.

The effects of WL, resulting from changes in AT, on catch decrease were important in 2 of these 16 lakes (Fig. 4f), where warmer AT were associated with lower WL which, in turn, resulted in lower catches. The effects of CHL, resulting from changes in AT, on catch decrease were not important in any of these 16 lakes. For the remaining 15 lakes, a mixed relationship was revealed as a simu- lated 25% catch decrease was associated with both colder and warmer AT.

A simulated 25% catch decrease was associated with both higher and lower PRE across the study lakes (Fig. 5a), but these effects were mediated through changes in both WL (Fig. 5d, f) and CHL (Fig. 5e, f). For 10 lakes, a simulated 25% catch decrease was associated with lower PRE (i.e., a positive relationship), with medians of 15–71% lower than the observed 1970–2014 medians (Fig. 5a). For another six lakes, a simulated 25% catch decrease was associated with higher PRE (i.e., a negative relationship), with medians of 20–36% higher than the observed 1970–2014 medians. For these 16 lakes, the simulated relationship between PRE and WL was consistently positive (Fig. 5b), while the simulated relationship between PRE and CHL could be either positive, negative, or mixed (Fig. 5c). Translating these simulated changes to effects on CATCHs revealed that the effects of CHL were important in 7 of these 16 lakes (i.e., leading to a >13.4%

catch decrease in simulations), but inconsistent relationships underlie this pattern: lower PRE was associated with both lower and higher CHL. The effects of WL, resulting from changes in PRE, on catch decrease were important only in 2 of these 16 lakes (Fig. 5d), where higher PRE increased WL. For the remaining 15 lakes, a mixed relationship was revealed as a simulated 25% catch decrease was associated with both lower and higher PRE.

Similarly, a simulated 25% catch decrease was associated with both increasing and decreasing LUag across the 31 study lakes (Fig. 6a), but effects of LUag on CATCHs were mostly mediated through changes in CHL (Fig. 6e, f) rather than through WL (Fig. 6d, f). For nine lakes, a simulated 25% catch decrease was

Category Symbol Description (unit)

Parameter and coefficient α Empirical coefficient that was not estimated

β BNM coefficient, which might be expressed a function ofαcoefficients μ Mean or predicted value

σ Standard deviation Variables associated with climate and land use AT Air temperature (°C)

LUag Agricultural land use (% of catchment area) LUun Undeveloped land use (% of catchment area) PE Potential evaporation (m)

PRE Precipitation (m) RUNOFF Catchment runoff (m³) Variables associated with lake environment ΔWL Change in water level (m)

CHL Concentration of chlorophylla(μg/l) WT Water temperature (°C)

Variables associated withfish catch,fish stocking, andfishing effort

B Fish biomass (kg/ha) CATCH Fish catch (kg/ha)

EFF Fishing effort (dimensionless) Q Fishing catchability (dimensionless)

ST Number offish stocked (number per hectare)

BNMBayesian networks model.

Table 2 Predicted effects between variable pairs in the BNM across lakes.

BNM variable pair Number of lakes Predictor Response Positive

effects

Negative effcts Mixed effects Air

temperature

Water temperature

31 0 0

Water level 6 17 8

Precipitation Water level 29 0 2

Chlorophylla 9 7 15

Agricultural land use

Water level 8 12 11

Chlorophylla 16 8 7

Water temperature

Chlorophylla 6 15 10

Fish catch 10 10 11

Water level Chlorophylla 13 8 10

Fish catch 9 5 17

Chlorophylla Fish catch 12 11 8

Fish stocking Fish catch 4 1 3

Fishing effort Fish catch 30 0 1

Effects between a variable pair were positive, negative, and mixed based on one-tailed 75%

credible intervals of BNM coefficients, for which complete summary statistics are given in Supplementary Data 4. Effects of stocking onfish catch were only included in eight lakes where stocked species contributed >20% offish catches in >10 years in the study period 1970–2014.

BNMBayesian networks model.

associated with decreasing LUag (i.e., a positive relationship), with median decreases of 3.0–6.9% of catchment area from the observed 1970–2014 medians (Fig. 6a). In turn, these simulated decreases in LUag resulted in lower CHL for eight of these nine lakes, but a higher CHL in the other (Fig. 6c). For another 12 lakes, a simulated 25% catch decrease was associated with increasing LUag (i.e., a negative relationship), with median increases of 1.2–14.0% of catchment area from the observed 1970–2014 medians (Fig. 6a). However, the simulated effects of increasing LUag on CHL were more variable across lakes, some associated with higher CHL and some associated with lower CHL (Fig. 6c). Translating these simulated changes to effects on CATCHs revealed that the effects of CHL, resulting from changes in LUag, were important in all 21 lakes (i.e., leading to a >13.4%

catch decrease), but once again the direction of the relationships

varied across lakes, particularly when LUag increased. The effects of WL, resulting from changes in LUag, on catch decrease were unimportant in these 21 lakes. For the remaining 10 lakes, a mixed relationship was revealed as a simulated 25% catch decrease was associated with both lower and higher LUag.

Characteristics of lakes where CATCHs are vulnerable. Our results showed that access to clean water (as measured by the proportion of population in the catchment using drinking-water and sanitation services

) was positively correlated with the magnitude of changes in each of AT, PRE, and LUag that was associated with a 25% catch decrease across the study lakes (AT:

r = 0.56, t-test p < 0.01, N = 31; PRE: r = 0.51, t-test p < 0.01, N = 31; LUag: r = 0.50, t-test p < 0.01, N = 31; Fig. 7a–c). This

Constance Kinneret Simcoe Geneva Taihu Maggiore Turkana Tanganyika Vänern Vättern Winnipeg Erie Okeechobee Naivasha Chilwa Titicaca Chapala Dongting Peipsi Issyk Kul Victoria Chaohu Balaton Tonle Sap Malawi Mweru Oneida Kariba Michigan Biwa Ladoga

–4 –2 0

Change in AT (°C) Africa Americas Asia Europe

2 4

Constance Kinneret Simcoe Geneva Taihu Maggiore Turkana Tanganyika Vänern Vättern Winnipeg Erie Okeechobee Naivasha Chilwa Titicaca Chapala Dongting Peipsi Issyk Kul Victoria Chaohu Balaton Tonle Sap Malawi Mweru Oneida Kariba Michigan Biwa Ladoga

–4 –2 0

Effects of AT on WT (°C)

2 4 –2 –1

Effects of AT on WL (m)

0 1 –100 0

Effects of AT on CHL (%)

100 200

–30 –20 –10

Effects of WT on fish catch (%)

0 10 –30 –20 –10

Effects of WL on fish catch (%)

0 10 –30 –20 –10

Effects of CHL on fish catch (%)

Summary

0 10

Air temperature

Water

temperature Water level

Chlorophyll a

Fish catch

a b c

e f g h

d

Fig. 4 Assessment for effects of air temperature onfish catches across the 31 study lakes.Changes or effects were calculated as differences (actual or relative) from the 1970–2014 medians.ashows simulated changes in air temperatures (AT) associated with a 25% decrease infish catches.b–dshow simulated effects of changes in AT on water temperature (WT), water level (WL), and chlorophylla(CHL).e–gdepict how simulated effects of WT, WL, and CHL, in response to changing AT, affectedfish catches.hsummarizes these linkages within a subset of the Bayesian networks model that demonstrates the effects of AT, where the width of each arrow is a qualitative indicator of the strength of the hypothesized causal effects. Ina–g, bars represent simulated median changes and error bars representedfirst and third quartiles of the simulated changes. Dashed lines divide the lakes into

“positive”(top 8 lakes),“mixed”(middle 15 lakes), or“negative”(bottom 8 lakes) associations between AT andfish catches, based on one-tailed 75%

credible intervals, as indicated by error bars ina. For positive lakes, a 25% decrease infish catches was associated with colder AT. For mixed lakes, a 25%

decrease infish catches was associated with both colder and warmer AT. For negative lakes, a 25% decrease infish catches was associated with warmer AT. Source data fora–gare provided as a Source Datafile.

Constance Kinneret

Simcoe

Geneva Taihu Maggiore

Turkana Tanganyika

Vänern Vättern Winnipeg

Erie Okeechobee

Naivasha Chilwa

Titicaca Chapala Dongting

Peipsi

Issyk Kul Victoria

Chaohu Balaton

Tonle Sap

Malawi Mweru Oneida Kariba

Michigan

Biwa Ladoga

Constance Kinneret

Simcoe

Geneva Taihu Maggiore

Turkana Tanganyika

Vänern Vättern Winnipeg

Erie Okeechobee

Naivasha Chilwa

Titicaca Chapala Dongting

Peipsi

Issyk Kul Victoria

Chaohu Balaton

Tonle Sap

Malawi Mweru Oneida Kariba

Michigan

Biwa Ladoga

–80 –40

Change in PRE (%)

0 40 –2 –1

Effects of PRE on WL (m)

0 1 –100 0

Effects of PRE on CHL (%)

100 200

–30 –20 –10

Effects of WL on fish catch (%)

0 10 –30 –20 –10

Effects of CHL on fish catch (%)

0 10 Summary

Precipitation

Water level

Chlorophyll a

Fish catch

a b c

e f

d

Africa Americas Asia Europe

Fig. 5 Assessment for effects of precipitation onfish catches across the 31 study lakes.Changes or effects were calculated as differences (actual or relative) from the 1970–2014 medians.ashows simulated changes in precipitation (PRE) associated with a 25% decrease infish catches.b,cshow simulated effects of changes in PRE on water level (WL) and chlorophylla(CHL).d,edepict how simulated effects of WL and CHL, in response to changing PRE, affectedfish catches.hsummarizes these linkages within a subset of the Bayesian networks model that demonstrates the effects of PRE, where the width of each arrow is a qualitative indicator of the strength of the hypothesized causal effects. Ina–e, bars represent simulated median changes and error bars representedfirst and third quartiles of the simulated changes. Dashed lines divide the lakes into“positive”(top 10 lakes),“mixed”(middle 15 lakes), or“negative”(bottom 6 lakes) associations between PRE andfish catches, based on one-tailed 75% credible intervals, as indicated by error bars in a. For positive lakes, a 25% decrease infish catches was associated with lower PRE. For mixed lakes, a 25% decrease infish catches was associated with both lower and higher PRE. For negative lakes, a 25% decrease infish catches was associated with higher PRE. Source data fora–eare provided as a Source Datafile.

Constance Kinneret

Simcoe

Geneva Taihu

Maggiore Turkana Tanganyika

Vänern Vättern

Winnipeg Erie Okeechobee

Naivasha

Chilwa

Titicaca Chapala Dongting

Peipsi

Issyk Kul Victoria

Chaohu Balaton Tonle Sap Malawi Mweru Oneida

Kariba

Michigan Biwa

Ladoga

Constance Kinneret

Simcoe

Geneva Taihu

Maggiore Turkana Tanganyika

Vänern Vättern

Winnipeg Erie Okeechobee

Naivasha

Chilwa

Titicaca Chapala Dongting

Peipsi

Issyk Kul Victoria

Chaohu Balaton Tonle Sap Malawi Mweru Oneida

Kariba

Michigan Biwa

Ladoga

–10 –5

Change in LUag (% of Acat)

0 5 10 15 –2 –1

Effects of LUag on WL (m)

0 1 –100 0

Effects of LUag on CHL (%)

100 200

–30 –20 –10

Effects of WL on fish catch (%)

0 10 –30 –20 –10

Effects of CHL on fish catch (%)

0 10 Summary

Agricultural land use

Water level

Chlorophyll a

Fish catch

a b c

e f

d

Africa Americas Asia Europe

Fig. 6 Assessment for effects of agricultural land use onfish catches across the 31 study lakes.Changes or effects were calculated as differences (actual or relative) from the 1970–2014 medians.ashows simulated changes in agricultural land use (LUag), in terms of percent of catchment area (Acat), associated with a 25% decrease infish catches.b,cshow simulated effects of changes in LUag on water level (WL) and chlorophylla(CHL).d,edepict how simulated effects of WL and CHL, in response to changing LUag, affectedfish catches.hsummarizes these linkages within a subset of the Bayesian networks model that demonstrates the effects of LUag, where the width of each arrow is a qualitative indicator of the strength of the hypothesized causal effects. Ina–e, bars represent simulated median changes and error bars representedfirst and third quartiles of the simulated changes. Dashed lines divide the lakes into“positive”(top 9 lakes),“mixed”(middle 10 lakes), or“negative”(bottom 12 lakes) associations between LUag andfish catches, based on one-tailed 75% credible intervals, as indicated by error bars ina. For positive lakes, a 25% decrease infish catches was associated with decreasing LUag.

For mixed lakes, a 25% decrease infish catches was associated with both decreasing and increasing LUag. For negative lakes, a 25% decrease infish catches was associated with increasing LUag. Source data fora–eare provided as a Source Datafile.

suggests that limited access to clean water is a characteristic for lakes where CATCHs are vulnerable to both climate and land-use changes, as relatively small changes (either increases or decreases) in AT, PRE, or LUag can cause a 25% catch decrease. Our results also showed that the magnitude of changes in LUag associated with a 25% catch decrease was negatively correlated with the shoreline development index (SDI), a measure of the circularity of lake surface and an indicator of littoral area relative to lake area

, across the study lakes (r = −0.39, t-test p = 0.03, N = 31; Fig. 7l).

This suggests that lakes with larger littoral area (as indicated by a higher SDI) were more vulnerable because they required relatively small changes in LUag (either increases or decreases) to cause a 25% catch decrease. Across the study lakes, however, there was no significant correlation (i.e., t-test p > 0.05, N = 31) for all com- binations between the magnitude of changes in each of AT, PRE, and LUag associated with a 25% catch decrease and shoreline population density or average depth (Fig. 7d–i), and between the magnitude of changes in each of AT and PRE associated with a 25% catch decrease and SDI (Fig. 7j, k).

Discussion

By using a Bayesian networks modeling approach to analyze time-series data for 31 lakes across five continents in the period 1970–2014, we found that climate and land-use changes could have both positive and negative effects on CATCHs. At one level, we found that effects of a climate or land-use driver (one of AT, PRE, and LUag in this study) on lake-environmental factors (WT,

WL, and primary productivity) were relatively consistent in directions (e.g., warmer AT always increased WT). However, the effects of a changing lake-environmental factor (e.g., CHL) could have either positive or negative effects on CATCHs across lakes (Table 2).

One likely explanation for the lack of a consistent response across lakes is the abiotic and biotic diversity of lake ecosystems that could not be incorporated within our global model. For example, increases in WT can have both positive and negative effects on fish growth and recruitment

, but the direction of these effects on individual fish species usually depends upon the thermal tolerance of this species and the thermal environment of the lake

. Within a lake, as most fisheries are multi-species and multi-gear in nature, fishers are able to adapt, usually gra- dually, to changes in the composition of fish species

. There is a rich literature reporting how different fish species responded differently to lake-environmental changes that could be linked to changes in climate and land use

, especially in lakes where some species are close to the boundary of their current distributions

. However, in response to climate and land-use changes, lake environment

may change at a very rapid pace so that native species cannot adapt, which, at an extreme, could create conditions for highly productive, introduced/exotic species to flourish and become dominant in CATCHs

. For lakes where the catch and effort could be subdivided into trophic levels or habitat guilds, future research could modify our existing modeling framework to predict catch at relevant trophic levels,

4

a

Africa Asia

Americas Europe

2 0 –2

Change in AT (°C)

–4

40 0 –40 –80 15

–5 0

0 25 50

Access to clean water (% of catchment population)

75 100

5 10

Change in Pann (%)Change in LUag (% of Acat)

4 16 64

Shoreline population (per km²)

256 1024 2 8 32

Average depth (m)

128 512 1 3 5

Shoreline development index 7 r = 0.56 (t-test p < 0.01)

r = 0.51 (t-test p < 0.01)

r = 0.50 (t-test p < 0.01) r = –0.39 (t-test p < 0.03)

d g j

b e h k

c f i l

Fig. 7 Identification of the characteristics of the vulnerable lakes.Thefigure shows simulated median magnitudes of changes, from the 1970–2014 medians, for air temperature (AT), precipitation (PRE), and agricultural land use (LUag, in terms of percent of catchment area (Acat)) associated with a 25% decrease, also from the 1970–2014 medians, infish catches and their associations with access to clean water (a–c), shoreline population density (d–f), average depth (g–i), and shoreline development index (SDI;j–l) across the 31 study lakes. Note that points in the negative section have negative simulated median changes associated with a 25% decrease infish catches. A lake with higher vulnerability offish catches to climate and land-use changes is indicated by lower magnitudes of changes in AT, PRE, and LUag associated with a 25% decrease, from the 1970–2014 medians, infish catches. A least- squares line is added when the correlation is significant atp=0.05 level (ttest,N=31). The access to clean water is measured by the proportion of population in the catchment using drinking-water and sanitation services. The shoreline population density is the population density within 10 km of a lake’s shoreline. SDI is a dimensionless index for circularity of the lake surface and a larger SDI indicates that the lake surface is less circular and may have a larger littoral area relative to lake area²⁴. Source data for thisfigure are provided as a Source Datafile.

for example, to evaluate whether a consistent catch response across lakes can be detected.

By assessing changes in climate and land-use drivers associated with a substantial, 25%, decrease in CATCHs across the 31 study lakes, we identified key processes that could underlie declines in CATCHs. Specifically, we found that changes in AT influence catches more through changes in WT than through changes in WL or primary productivity (as measured by CHL; Fig. 4h). Our BNM revealed strong positive relationships between AT and WT across all the 31 study lakes, which is consistent with predictions based on energy balance and findings in another global study

. Given robust projections for climate warming throughout this century

, our results suggest that projected increases in AT will result in increases in WT, but will not necessarily lead to wide- spread declines in CATCHs across lakes, as our simulations revealed that decreases in CATCHs were associated with warmer AT in only 7 of the 31 study lakes. One important caveat is that the largest AT increase in our model was 2.6 °C warmer than the 1970–2014 median. We caution against extrapolating our results should AT warm to even higher temperatures.

We found that changes in PRE could lead to substantial declines in CATCHs through changes in either WL or primary productivity (as measured by CHL; Fig. 5f). Across 29 of the 31 study lakes, our BNM suggested strong positive relationships between PRE and WL, which is consistent with water cycle and runoff dynamics. However, the lack of any consistent pattern of linkages between PRE and primary productivity revealed in our BNM exemplifies the diversity of lake systems. In theory, increased PRE can result in increases in primary productivity in a lake, directly through increasing inputs of allochthonous-derived nutrients from the catchment

and indirectly through increasing WL that lead to the release of nutrients from inundated riparian land

. Nonetheless, depending on the timing, duration, and intensity of PRE, the increased PRE can also result in increases in sediment loads and resuspension, which can result in decreases in water clarity and primary productivity

. Based on current pro- jections for this century, generally, future PRE may increase in the high-latitude (>40°) and near-equator (latitude <10°) regions, but decrease in mid-latitude regions

. As shown in Supplementary Data 2, these projections suggest that 21 of the 31 study lakes will have PRE increases, yet the scenarios under which these 21 lakes experienced decreases in CATCHs in our simulations included increased PRE (four lakes), reduced PRE (eight lakes), and mixed (nine lakes). While these projections suggest that 10 of the 31 study lakes will have PRE decreases (Supplementary Data 2), the scenarios under which these 10 lakes experienced decreases in CATCHs in our simulations also included increased PRE (two lakes), reduced PRE (two lakes), and mixed (six lakes). As a result, although projected changes in PRE will result in consistent directional changes in WL, it will not necessarily result in the same directional changes in primary production and CATCHs across lakes. Finally, although the projected changes in PRE across global change scenarios

are generally within the range of our BNM inputs (i.e., from 63% lower to 51% higher than the 1970–2014 median), we caution against extrapolating our results should PRE change beyond this range.

While increases in LUag could affect lake ecosystems increas- ing water withdrawals and nutrient inputs, which could result in decreases in WL and increases in primary productivity

, we found that changes in LUag affected CATCHs more through changes in primary productivity (as measured by CHL) than through changes in WL (Fig. 6f). However, our BNM showed that increases in LUag could result in increases in primary pro- ductivity in 17 of the 31 study lakes, which suggests that in the remaining 14 study lakes, the positive relationship between nutrient inputs and LUag was decoupled. One plausible

explanation for this decoupling is that the intensity of agriculture in a lake’s catchment could increase with little or no change in the area of agricultural land, such as through increases in fertilizer use

, livestock densities

, or cage culture

. Additionally, the area of LUag does not reflect the management efforts to control nutrient enrichment in recipient lakes, which is very common in developed countries

, but less so in developing countries.

Generally, LUag is projected to increase in developing coun- tries but decrease in developed countries in this century

. As shown in Supplementary Data 2, 16 of the 31 study lakes are projected to have decreases in LUag and the scenarios under which these lakes experienced decreases in CATCHs in our simulations included increased LUag (eight lakes), reduced LUag (two lakes), and mixed (six lakes). Similarly, among 15 lakes that are projected to undergo increases in LUag, the scenarios under which these lakes experienced decreases in CATCHs in our simulations included increased LUag (four lakes), reduced LUag (seven lakes), and mixed (four lakes). Once again, our results suggest that projected changes in LUag will not necessarily result in the same directional changes in primary production and CATCHs across lakes and we caution against extrapolating our results should agricultural land-use change beyond the range of our BNM inputs (from an decrease of 13% to an increase of 11%, in terms of percent of catchment area, from the 1970–2014 median).

We identified low access to clean water (as measured by the proportion of population using drinking-water and sanitation services in the catchment) as one characteristic of lakes where CATCHs are vulnerable to changes in AT, PRE, and LUag (Figs. 7a–c and 8). Access to clean water can only be improved with substantial investments for water-use efficiency and water- quality protection

. Therefore, combining with the identified key processes that underlie changes in CATCHs (Figs. 4h, 5f, and 6f), these results suggest that the investments for water-use efficiency and water-quality protection can also make CATCHs less vul- nerable to climate and land-use changes. Future research will be required to determine whether this correlation is causative; one potential mechanism that has support from previous studies is that investments in sanitation improve fish habitat

. One caveat is that lakes in regions with more access to clean water may also have stronger fishing regulations, which could also make CATCHs less vulnerable to climate and land-use changes by sustaining healthier fish stocks. We were unable to disentangle these effects in our analysis because none of the 31 study lakes has both high access to clean water and limited or weak fishing regulations.

Independent of access to clean water, our results suggest that climate and land-use changes can result in substantial decreases in CATCHs across lakes with very different socio-economic and hydrogeomorphological characteristics, such as shoreline popu- lation densities, depths, and littoral areas (Figs. 7d–k and 8).

Although we identified a large littoral area (as indicated by SDI) as one characteristic for lakes where CATCHs are vulnerable to changes in LUag, the significant relationship was driven by two lakes with SDI >7 (Fig. 7l). Therefore, more research is needed to confirm this suggestion.

Throughout this study, we interpreted the results across the 31 study lakes collectively, with a focus on understanding effects of climate and land-use on CATCHs at a global scale (Figs. 3–7).

We caution against interpreting our lake-specific results because

the uncertainty around some predictions can be large, and it is

difficult to discern whether this is a consequence of structural

errors in our BNM or the uncertainty that arises from using data

from so many sources (Supplementary Data 3). However, we

suggest using our lake-specific predictions as priors for future

lake-specific assessments, where quantitative models can be

tailored to include important, lake-specific drivers that we were not able to include in our global model and data collected by consistent methods can be obtained. This is especially relevant to lakes where effects of climate or land-use change alone on CATCHs has not been as prevalent as effects of other anthro- pogenic stressors, such as pollution, invasive species, and over- fishing; however, climate and land-use changes may exacerbate the negative effects of these stressors on CATCHs

.

In conclusion, although we were unable to identify consistent directional relationships between lake-environmental variables (e.g., WT, WL, CHL) and CATCH (Table 2; Figs. 4–6), correla- tions based on our simulations suggested an intriguing and potentially important relationship between access to clean water and vulnerability to a large reduction in CATCH in response to climate and land-use changes (Figs. 7 and 8). As an extension, our results suggest that CATCHs for lakes in developing countries in sub-Saharan Africa, Southeastern and Central Asia, and Central and South America are more vulnerable, which is consequential given that inland CATCHs are among the highest in the world in these regions and that the main purpose of fishing is for food rather than income

. To mitigate the threat of significant decreases in CATCHs and food security, our results imply that development strategies that include investments for clean water

can also benefit lake CATCHs in the face of anthropogenic stressors (Fig. 8). The implications of this result made possible by analyzing time series of lakes across continents highlight possible synergies for policy makers aiming to achieve multiple United Nation Sustainable Development Goals (SDGs)

, including Zero Hunger (SDG 2), Clean Water and Sanitation (SDG 6), Climate

Action (SDG 13), and Life Below Water (SDG 14), and Life on Land (SDG 15).

Methods

Choice of modeling approach. We chose to use a Bayesian networks modeling approach because it could minimize negative effects of data limitations on the statistical power of our analysis, by accounting for our understanding of multi- level, causal processes that could underlie changes in CATCHs in response to changes in climate and land use across lakes. A BNM could be developed to represent the hypothesized causal processes by expressing them as conditional probabilistic relationships between model variables, with which the estimation of every BNM coefficient took all model inputs derived from data into account²⁰. For example, if we developed a simple BNM to represent that changes in AT resulted in changes in WT, which, in turn, resulted in changes in CATCH in a lake, the estimation of model coefficients for the relationship between AT and WT and for the relationship between WT and CATCH would both take model inputs for all variables (AT, WT, and CATCH) into account. Although the 31 study lakes were selected partly because of good data availability, there were still large within-lake differences in data availability among model variables (as shown in Supplementary Data 3). For example, we obtained 9, 16, and at least 32 years of data for CHL, WT, and all the other variables in Lake Chilwa, respectively. Therefore, the BNM’s capability of taking all model inputs into account for the estimation of every model coefficient was important to maximize the statistical power of our analysis.

Model development. To mathematically formulate our BNM based on our hypothesized causal processes (Fig.2), we assumed that (1) each response variable in the BNM has a normal distribution, with a mean that can be expressed as a function of predictor variables based on the primary literature described below, and (2) annual time-series data are independent observations, such that the mean value of each response variable in 1 year is dependent upon the values of predictor variables in the same year, but independent from all of its mean values in the other years. For example, the conditional probability distribution of WT in our Vulnerability to climate and land-use changes

Low

Africa Catch (kg ha^–1)

Americas 2

25 125 Asia

Europe Vänern

Dongting

High

Vättern

Balaton

Ladoga

Peipsi Maggiore

Michigan Erie ConstanceBiwa

Geneva Oneida Okeechobee

Chapala Kinneret

Chaohu Taihu Issyk Kul Winnipeg

Titicaca Simcoe

Kariba Tonle Sap

Victoria

Turkana

Malawi Mweru Naivasha Chilwa Tanganyika

Vulnerability to land-use changes

LowHigh

100 7

5

Shoreline development index

3

1

75

Access to clean water (% of catchment population)

50 25 0

Fig. 8 Distribution of the 31 study lakes with respect to access to clean water and shoreline development index.The area of each circle is proportional to the medianfish catch in the period 1970–2014. Based on the correlation analysis shown in Fig.7, lakes located in areas with lower access to clean water are more vulnerable to substantial decreases infish catches driven by either climate or land-use change and lakes with larger shoreline development index are more vulnerable to substantial decreases infish catches driven by land-use change. Source data for thisfigure are provided as a Source Datafile.

BNM was expressed as

WT_iNORMμ^WT_i ;σ^WT

; μ^WTi ¼β^WT0;i þβ^WT1;i ´AT_i; (

ð1Þ where NORM(μ,σ) represents a normal distribution with a mean ofμand a standard deviation ofσ, subscriptediindicates theith lake, superscripted WT indicates that the parameter or BNM coefficient was associated with the response variable WT, andβrepresents a BNM coefficient. In Eq. (1),μ^WTwas expressed as a linear function of AT based on the primary literature²¹and was assumed to be dependent upon the observed value of AT in the same year, but independent from anyμ^WTof the other years. In a similar manner, we give details for how we developed the conditional probability distributions of the other three BNM response variables—change in water level (ΔWL), CHL, and CATCH—in the following of this section.

The conditional probability distribution ofΔWL in our BNM was developed based on an empirical linear relationship betweenΔWL and annual catchment runoff (RUNOFF)²³, of which the latter could be expressed as a balance between PRE and evapotranspiration (ET)⁵³. In a simple manner, these relationships could be expressed as

ΔWL¼α^ΔWL0 þα^ΔWL1 ´RUNOFF; ð2Þ RUNOFF¼PREα^ET1 ´PE´LUagα^ET2 ´PE´LUun; ð3Þ whereαrepresents an empirical coefficient, PE is potential evaporation, LUun is undeveloped land use, andα1ET× PE × LUag andα2ET× PE × LUun represent ET over agricultural area and undeveloped area, respectively. In this study, we categorized catchment land use into three types: agricultural (which includes cropland, managed pasture, and rangeland), undeveloped (which includes forest, primary and secondary non-forest land, the lake itself, and other water and ice area), and urban. Equation (3) did not include urban land use because urban water use is mostly non-consumptive and returns to the catchment⁵⁴. By combining Eqs.

(2) and (3) and re-arrangingαcoefficients, the conditional probability distribution ofΔWL in our BNM was expressed as

ΔWLiNORMμ^ΔWLi ;σ^ΔWL

;

μ^ΔWL_i ¼β^ΔWL_0;i þβ^ΔWL_1;i ´PREiþβ^ΔWL_2;i ´PEiþβ^ΔWL_3;i ´PEi´LUag_i; (

ð4Þ where eachβcoefficient can be expressed as a combination ofαcoefficients in Eqs.

(2) and (3). We only included LUag in Eq. (4) because in each of these 31 study lakes, LUag+LUun in the catchment was almost a constant, which means that LUun could be replaced with a constant minus LUag, as the increase in agricultural area mostly resulted from the decrease in undeveloped area, or vice versa, in the study period.

The conditional probability distribution of CHL in our BNM was developed based on empirical log–log relationships between CHL and total phosphorus²⁴and between CHL and WT²⁵. Because total phosphorus data were only available for a few of our study lakes, we modeled the total phosphorus in lakes as a function of PRE²⁶, LUag²⁶, andΔWL²⁹. This led to a conditional probabilistic distribution for CHL expressed as

logðCHLiÞ NORMμ^CHL_i ;σ^CHL

;

μ^CHLi ¼β^CHL_0;i þβ^CHL_1;i ´logðPREiÞ þβ^CHL_2;i ´logðLUagiÞ þβ^CHL_3;i ´μ^ΔWLi þβ^CHL_4;i ´μ^WTi ; (

ð5Þ whereμ^WTandμ^ΔWLare from Eqs. (1) and (4), respectively.

Finally, the conditional probability distribution of CATCH was developed based on (1) a theoretical relationship between CATCH andfish biomass (B)³⁴, (2) a hypothesized empirical relationship betweenfishing catchability (Q) and EFF and B³⁴, and (3) a hypothesized empirical relationship betweenBand WT^27,28, ΔWL^29,30, and CHL^31,32and number offish stocked (ST)^35,36. In a simple manner, these relationships could be expressed as

CATCH¼Q´EFF´B)logðCATCHÞ ¼logðQÞ þlogðEFFÞ þlogðBÞ; ð6Þ logðQÞ ¼α^Q0þα^Q1´logðEFFÞ þα^Q2´logðBÞ; ð7Þ logðBÞ ¼α^B0þα^B1´WTþα^B2´ΔWLþα^B3´logðCHLÞ þα^B4´logðSTÞ; ð8Þ By combining Eqs. (6)–(8) and re-arrangingαcoefficients, we developed the conditional probabilistic distribution for CATCH expressed as

logðCATCH_iÞ NORMμ^CATCHi ;σ^CATCH

;

μ^CATCHi ¼β^CATCH0;i þβ^CATCH1;i ´μ^WTi þβ^CATCH2;i ´μ^ΔWLi þβ^CATCH3;i ´μ^CHLi

þβ^CATCH_4;i ´logðSTiÞ þβ^CATCH_5;i ´logðEFFiÞ;

8>

><

>>

: ð9Þ

whereμ^WT,μ^ΔWL, andμ^CHLare from Eqs. (1), (4), and (5), respectively. As described in the Supplementary Methods, we did not estimateβ4,iCATCHfor 23 lakes where stocking effects on CATCHs were not sufficiently important in the study period.

The derivations and data sources of model inputs, including climate and land- use drivers (AT, PRE, and LUag), lake-environmental factors (WT, WL, and CHL), CATCH, ST, and EFF are given in the Supplementary Methods.

Prior distributions. We followed a previously used procedure²⁰to derive prior distributions of BNM coefficients. For example, to derive the prior distribution of β^WTi , wefirst expressed the function ofμ^WTi in Eq. (1) as 31 linear models (i.e., one for each study lake) in a form

WT_i¼β^WT_0;i þβ^WT_1;i ´AT_iþε^WT_i ; ð10Þ whereεis the residual error. Then we derive the least-squares estimates for the coefficients of each of these 31 linear models based on model inputs derived from data for WTiand ATiin the same year. In theory, the least-squares estimates for the coefficients of each of these 31 linear models have a multivariate normal (MVN) distribution because they could be expressed as linear combinations of WTi, which was assumed to have a normal distribution as shown in Eq. (1). Finally, we used the MVN distribution for the least-squares estimates of the linear model coefficients as the prior distribution of each of the 31 sets ofβiWT, but the cov- ariance matrix of the MVN distribution was set to be proportional to its un-scaled, least-squares estimates, with a proportional constant (i.e., the hyperparameter) having a noninformative (hyperprior) distribution²⁰. We repeated this procedure to derive prior distributions forβiΔWL,βiCHL, andβiCATCH. The prior distributions of these BNM coefficients were always developed on a lake-by-lake basis because for each of the 31 study lakes, we were able to obtain data to derive model inputs for at least 9 years for all BNM variables (refer to Supplementary Data 3 for data sources and availability). Lastly, for nuisance parametersσ^WT,σ^ΔWL,σ^CHL, and σ^CATCHin Eqs. (1), (4), (5), and (9), respectively, we set each of them to have a noninformative prior distribution.

Coefficient estimation. The posterior distributions of our BNM coefficients were developed based on pooled information across the 31 study lakes. Our BNM was structured in a way to have the transferability of information across the lakes informed by data. Specifically, in the derivation of posterior distribu- tions of our BNM coefficients, the relative importance of lake-specific infor- mation increases with lake-specific sample size and across-lake difference. If a lake was very different from all the other lakes, for example, the relative importance of its lake-specific information would be much higher than across- lake information in the derivation of posterior distributions of lake-specific BNM coefficients.

Computationally, we derived empirical posterior distributions for all BNM coefficients by using a Gibbs sampling method to run Markov chain Monte Carlo (MCMC) simulations⁵⁵. We ran three random-starting MCMC chains by implementing the program JAGS⁵⁶and the package runjags⁵⁷in R⁵⁸. After discarding 20,000 iterations for burn-in and adaptation, we considered that the convergence of three MCMC chains was achieved at 50,000 interactions as autocorrelations were close to 0 and Gelman and Rubin’s convergence diagnostics⁵⁹were <1.005 for all coefficients. Our empirical posterior distribution of each BNM coefficient was based on 50,000 samples from each converged MCMC chain.

Effects between variable pairs. As shown in Fig.2, there were 13 pairs of pre- dictor and response variables in our BNM. For each of these 13 variable pairs, we categorized the effects of predictor variable on response variable as positive, negative, and mixed in each lake based on the posterior distribution of the asso- ciated BNM coefficient, which we used as a proxy of effect strength because it represents how much the response variable would change in response to one unit change in the predictor variable. For example, we categorized the effects of AT on WT in lakeibased on the posterior distribution ofβ1,iWT, which represents how much of WT would increase in response to one unit change of AT. We categorized effects between a variable pair as positive and negative based on the 75% CI of the associated BNM coefficient. We categorized effects between a variable pair as mixed when both positive and negative 75% CIs of the associated BNM coefficients included 0, which, statistically, also means that its 50% CI included 0.

Changes in drivers associated with a 25% catch decrease. We implemented the BNM and used a Monte Carlo simulation method to empirically derive distribu- tions of climate (AT and PRE) and land-use (LUag) drivers associated with a 25%

decrease in CATCH from its 1970–2014 median (i.e., a 25% catch decrease) in each of the 31 study lakes. We used a 25% catch decrease as our simulation target because it is close to the maximum value of standard errors for the ratios of CATCH to its median, which ranged from 2.1% to 25.7% across the study lakes in the study period.

Monte Carlo simulation. Our Monte Carlo simulation was a three-step process.

Thefirst step was to randomly sample a set of BNM coefficients (i.e.,β^WT,β^ΔWL, β^CHL, andβ^CATCH) from empirical posterior distributions by implementing the package rv⁶⁰in R. Then, we calculated predicted medians of WT,ΔWL, CHL, and